Improvised Explosive Devices (IEDs), in different forms, are a favorite weapon for global in- surgents and terrorists, with an average of 260 IED incidents per month in Afghanistan and Iraq for the time from January 2004 until May 2010. Terrorist’s ability to quickly change the IED type requires a flexible approach in IED detection training. To create an effective training tool, an image taken in the field must be segmented into two sections, separating the IED and the image background.
Different from objects typically used to evaluate the quality of segmentation algorithms, IEDs are not easy to detect even for humans. They are purposefully hard to detect, concealed or colored very similar to the background. This thesis evaluates existing algorithms for image segmentation and creates new algorithms to pre-process colored images in order to segment IEDs.
We first modify a basic graph cut algorithm to segment completely visible IEDs from a uni- formly colored background. To measure the performance depending on the parameter settings, we run a series of tests defined by a Nearly Orthogonal Latin Hypercube design of experiment. A second series of tests shows that the basic approach does not result in good segmentations if applied to realistic IEDs. We identify two major issues that are due the special nature if IEDs:
Different from other segmentation tasks, real-world IEDs are only a small fraction of the complete image. This results in a training area for the background model that is extremely large compare to the training area of the IED.
IEDs do not necessary appear as a single object. Often, charges and initiation systems appear -if visible- as different objects. Visible parts of an IEDs can be small and partially hidden. Therefore, a single, manually provided training area of an IED is not precise enough.
We address the increased difficulty while segmenting IEDs with three key contributions:
Our algorithm automatically divides a user-defined background area into smaller areas. We generate separate color models for each of these areas, to ensures that a color model includes only colors that appear close together in the background. For each of these areas, we build a color model that represents the complete RGB color space of the area.
We compress each of these complex color models into a statistical model by applying the k-means Clustering algorithm for up to 10 clusters. This compressed statistical descrip- tion increases the number of background models we can hold simultaneously in working memory.
We use the estimated distance to the background to estimate the initial object label auto- matically. In this way, we can segment an IED without user input labeling the IED.
The results of this thesis have three different applications: 1) defining the IED on an image, 2) generating new training images, 3) helping humans detect IEDs.
In order to distinguish between a student’s detection and misdetection of an IED in a colored image, the area of an IED must be defined. Our work defines this area through the set of pixels at the end of the graph cut algorithm.
To generate new training images, a segmented IED can be combined with images showing a different background. Our work facilitates this through precise IED segmentation.
Finally, the automated IED labels we generate may be helpful as a hybrid between a computer pre-processing the image, and a human operator providing the final identification. The auto- matic label computation is efficient and can be computed in real-time.